A customer sent a long prompt to Cloudflare Workers AI. The model answered with text from another customer’s request. There was no malicious instruction hidden in the prompt. The inference system read memory pages that belonged to a previous request.
Cloudflare disclosed in October that its GLM-5.3 deployment produced 155 affected responses between August 27 and September 9, 2026. All involved prompts longer than 681,280 tokens. Cloudflare says it found no evidence of deliberate exploitation and confirmed that no other Workers AI model was affected. The incident is a concrete example of a tenant boundary failing inside the machinery that serves a model.
TL;DR
- The GLM-5.3 configuration advertised a 1,048,576-token context window, while its prefill KV-cache pool held only 681,280 tokens.
- When an over-capacity request was rejected internally, an SGLang error-path bug allowed decode to continue with stale KV pages from earlier requests. The response could include another customer’s data.
- Cloudflare corrected the configuration and engine, restored the model on September 9, and later applied broader hardening. The relevant SGLang fixes reached releases through v0.5.22.
- Operators of disaggregated SGLang should check the running pool size against the advertised context length, patch the engine, test rejection paths and monitor for cross-request leakage.
Why This Matters
Most AI security reviews look at prompts, tools and access tokens. In a shared inference service, the engine also has to keep one tenant’s working state out of another tenant’s answer. The KV cache holds intermediate state used while generating text. It is reused for performance; safety depends on correct page ownership and lifetime tracking.
Cloudflare’s deployment split work across prefill and decode GPUs. Prefill processed the prompt and transferred its KV state to decode, which generated the response. The two sides had to agree on how many pages the request owned and which pages had actually been written. In the failed path, they did not.
How the Cross-Request Read Happened
The advertised context length exceeded the prefill pool’s capacity. A prompt above 681,280 tokens could therefore be valid according to the model’s published limit while still exceeding the available KV pool.
SGLang marked such a request for rejection and substituted a one-token placeholder. In the disaggregated path, that state did not stop the request at admission. Prefill computed and transferred state for the placeholder, then told decode that the transfer was complete. Decode had reserved pages for the original long prompt. It treated the completion signal as proof that all those pages contained this request’s data, although most had not been overwritten. Generation then attended to pages left by earlier requests.
The outcome could look deceptively normal at the HTTP layer: Cloudflare says the affected requests returned success and coherent output. Its pre-launch long-context tests checked status codes and streaming behavior, but did not catch the content mismatch. A successful response is therefore a poor validation for this failure mode.
This was a specific combination of configuration and engine behavior. It does not imply that every SGLang deployment or every model provider leaked data. It does show why a model’s advertised context window and the running engine’s memory allocation must be checked together.
What Operators Can Verify
For a disaggregated SGLang service, collect the deployed engine version, model configuration, advertised maximum context, and the actual KV-pool capacity reported at startup or through metrics. Check both prefill and decode roles. The capacity can change with model weights, quantization and other deployment settings, so a configuration-file review alone is insufficient.
Then test boundary cases in an isolated environment. Send prompts below, at and above the usable pool limit, including requests that the scheduler should reject. A rejected request should return an error without reaching decode. Use distinct synthetic canary strings in separate tenant sessions and inspect complete outputs for crossover. Do not place real customer data in these tests.
Cloudflare recommends SGLang v0.5.22 or backported fixes for operators using disaggregated inference. The incident-specific admission fixes appeared in earlier releases, while later releases included additional work on cancellations, page lifetime and transfer integrity. Version alone cannot establish that a deployment’s context and pool sizes agree; verify both.
For incident response, define the affected model and time window first. Preserve request identifiers, tenant mapping, token counts, engine deployment history and any retained outputs under appropriate access controls. A response containing unfamiliar content is a stronger lead than an HTTP error counter. If the provider operates the model, ask for its exposure assessment and whether it can identify affected responses; do not assume your own API logs contain generated text.
What You Can Do Today
- Identify shared inference deployments. Record which models use disaggregated prefill and decode, which tenants share workers, and who owns the engine version.
- Compare live limits. Measure the actual KV-pool capacity against each model’s context limit after every deployment change.
- Patch and exercise error paths. Use the relevant SGLang fixes, then test over-capacity requests, aborts and client disconnects with synthetic tenant canaries.
- Review telemetry. Alert on impossible combinations such as a very long prompt completing suspiciously quickly, and make it possible to trace a response to the model build and request without collecting prompt content unnecessarily.
- Prepare for disclosure. Decide in advance how to preserve and review potentially affected outputs while limiting access to customer data.
Related Posts
- LLMJacking: Detect Stolen AI Access Before the Bill Arrives — an identity-driven failure at a different AI service boundary.
- LiteLLM CVE-2026-59822: Your AI Gateway Is a Cloud Control Plane — why the gateway in front of a model needs its own security review.
Sources
Useful read?
Find us again on Google.
Add Hive Security as a preferred source for practical security research and analysis.
Add as a preferred source on GoogleChoose Hive Security in Google's source preferences.